Nuclear power sensor data reconstruction method and system based on Seq2Seq network model
By using a Seq2Seq network model and copula entropy analysis, a composite sensor array was constructed, which solved the data failure problem of nuclear power plant sensors under high temperature, high pressure and strong radiation environments, and achieved high-precision data reconstruction, thereby improving the safety and reliability of nuclear power plants.
Patent Information
- Application Number
- CN202511193131.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-25
- Publication Date
- 2025-12-16
AI Technical Summary
Existing technologies are insufficient to effectively address data failures caused by malfunctions of nuclear power plant sensors under high temperature, high pressure, and strong radiation environments, which affects the safety and stability of nuclear power plants.
A Seq2Seq network model combined with copula entropy analysis is used to construct a composite sensor array. Spatiotemporal features are captured by bidirectional LSTM to achieve high-precision reconstruction of sensor data. The reconstruction model includes an encoder and a decoder. Fault data repair of the target sensor is performed using the composite sensor data.
It enables high-precision reconstruction of sensor data under fault conditions, improves the safety and reliability of nuclear power plants, reduces the safety risks caused by single sensor failures, supports fault detection and location, and improves the system's fault tolerance.
Smart Images
Figure CN121145928A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of intelligent operation and maintenance of nuclear power sensors, and particularly relates to a nuclear power sensor data reconstruction method and system based on a Seq2Seq network model. BACKGROUND
[0002] Nuclear power plant sensors play a crucial role in ensuring the safe and stable operation of the reactor system. They are responsible for continuously monitoring key parameters such as temperature, pressure, flow rate, radiation dose, and water level. These sensor readings are not only essential for daily operation but also directly related to the safety and stability of the entire nuclear power plant. During reactor operation, sensor data helps operators monitor the status of the reactor and its subsystems in real-time, ensuring that core parameters remain within safe operating ranges. This helps to avoid dangerous situations such as overheating and overpressure, and enables early detection of potential faults or signs of accidents. However, the harsh operating environment of nuclear power plants, including high temperatures, high pressures, and strong radiation, greatly increases the probability of sensor failure. The high-temperature environment inside the reactor can cause degradation, corrosion, or complete failure of sensor materials; extreme pressure conditions can damage the sensor housing; and strong radiation can distort or permanently damage sensor signals. Once a sensor fails or its data becomes unavailable, its output reliability is lost, leading to invalid readings. This not only hinders operators from accurately determining system status, but also weakens the effectiveness of decision support systems, thereby increasing the risk of major safety accidents. Therefore, it is of great significance to study data reconstruction techniques for nuclear power plant sensors in fault states. Data reconstruction techniques can predict and compensate for missing or distorted sensor readings using information provided by other normal sensors, ensuring the integrity and accuracy of monitoring data and improving the safety of nuclear power plant operation. Such methods not only reduce the safety risks caused by single sensor failures, but also improve system fault tolerance, maintaining monitoring and control functions in extreme situations. In addition, reconstruction techniques can also support fault detection systems in identifying sensor abnormalities, providing precise fault location and recovery strategies for maintenance personnel, avoiding unnecessary equipment downtime, and improving the economic efficiency and operational reliability of nuclear power plants.
[0003] In recent years, deep learning techniques have made significant progress in time series prediction. Recurrent neural networks represented by LSTM and GRU have excellent time series modeling capabilities and have performed well in single-variable time series prediction tasks such as wind speed prediction and power load prediction. However, these methods mainly model single variables and are difficult to directly apply to the complex scenario of multi-sensor collaborative reconstruction. In addition, when using multiple sensors to reconstruct data for a target sensor, it is necessary to consider whether there is a certain correlation between them.
[0004] In summary, a nuclear power sensor data reconstruction method and system based on a Seq2Seq network model are developed, which is of great significance to ensure the safe and stable operation of nuclear power plants. SUMMARY
[0005] The purpose of the present application is a nuclear power sensor data reconstruction method and system based on a Seq2Seq network model, aiming to solve the shortcomings of existing nuclear power sensor data reconstruction methods in nonlinear modeling, long sequence dependence processing and multi-sensor collaborative utilization. By introducing an associated sensor reconstruction module, a high-precision and robust nuclear power sensor data reconstruction method is realized to solve the abnormal data repair (reconstruction) of nuclear power sensors under fault conditions and ensure the safety and reliability of nuclear power plant operation.
[0006] The purpose of the present application is achieved by the following technical solutions:
[0007] A data reconstruction method based on a Seq2Seq network model, comprising the following steps:
[0008] Step 1: Obtain steady-state normal operation data and fault data through a nuclear power plant simulation platform and preprocess them, and select a target sensor;
[0009] Step 2: Use copula entropy as a correlation measurement method to identify the sensor combination associated with the target sensor under steady-state conditions;
[0010] Step 3: Based on the correlation structure in step 2, use the associated sensor combination and target sensor data in the normal operation data to train and construct a Seq2Seq model;
[0011] Step 4: Use the associated sensor combination data in the fault data to output the reconstruction data of the target sensor through the trained Seq2Seq model, realizing the data reconstruction function of the target sensor under fault conditions.
[0012] Further, the sensor combination does not include the target sensor itself.
[0013] Further, the calculation formula of copula entropy in step 2 is:
[0014] H c (X,Y)=-∫ u c(u,v)logc(u,v)dudv
[0015] Where u=F X (x), v=F Y (y) are the edge cumulative distribution function transformations of N-dimensional random variables X and Y, and X and Y are used for target sensor and sensor data used for correlation calculation in the data set.
[0016] Further, after obtaining the copula entropy values of the target sensor and the remaining sensors, the correlation calculation results are presented in the form of a heat matrix chart, with darker colors representing stronger correlation and vice versa; then, the sensor combination with the strongest correlation with the target sensor is selected by the color of the elements in the heat map and the size of the entropy values, and the normal operation data of the sensor combination with the strongest correlation is used as the input of the subsequent data reconstruction model.
[0017] Further, the training data of the Seq2Seq model is preprocessed:
[0018] The input data is constructed with a sliding window of 10 time steps, covering the historical time t-9 to the current time t;
[0019] The time series data of each sensor is individually normalized by MinMax to the range [0, 1];
[0020] A three-dimensional input tensor with dimensions [batch size, time step, feature number] is generated.
[0021] Further, the Seq2Seq model constructed in step 3 includes:
[0022] Encoder: The input data is processed by a bidirectional LSTM layer, the forward LSTM layer captures the forward time dependence from left to right, and the reverse LSTM layer captures the reverse time dependence from right to left, each time step outputs a 128-dimensional hidden state after bidirectional splicing, and the hidden state at the final time is used as a 128-dimensional global context vector, carrying the compression information of the entire sequence;
[0023] Decoder: The context vector of the encoder is used as the initial state, and the reconstruction sequence of the target sensor is generated through the step-by-step autoregressive prediction mechanism, each step receives the previous prediction value and outputs a 1-dimensional prediction result through a fully connected layer, and the complete sequence is constructed by repeating 10 times.
[0024] Further, the output data reconstruction in step 4 is post-processed:
[0025] De-normalization operation: The 0-1 range prediction value output by the model is mapped back to the original physical dimension;
[0026] Sliding window integration: The overlapping window prediction results are fused by weighted averaging to eliminate boundary effects;
[0027] Final output: The target sensor reconstruction value is generated, which is completely synchronized with the original data sampling rate.
[0028] Further, the training termination condition of the Seq2Seq model is: reaching the preset maximum number of iterations or the mean square error of the model output and the theoretical output meeting the convergence threshold.
[0029] A computer device / equipment / system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of a nuclear power plant sensor data reconstruction method based on a Seq2Seq network model.
[0030] A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of a nuclear power plant sensor data reconstruction method based on a Seq2Seq network model.
[0031] The beneficial effects of this invention are as follows:
[0032] This invention presents a nuclear power plant sensor data reconstruction method based on a Seq2Seq network model, which offers significant advantages: It employs a deep learning-based Seq2Seq architecture to overcome the limitations of traditional linear methods. Through copula entropy correlation analysis, it establishes a sensor combination (excluding the target sensor itself) associated with the target sensor, and uses this sensor combination to reconstruct the target sensor's readings under fault conditions. Furthermore, the model incorporates a bidirectional LSTM to capture spatiotemporal features, achieving high-precision end-to-end modeling. Attached Figure Description
[0033] Figure 1 Flowchart for Seq2Seq model training and testing;
[0034] Figure 2 Here is a diagram of the Seq2Seq model structure;
[0035] Figure 3 This is a framework diagram of a nuclear power plant sensor data reconstruction method based on a Seq2Seq network model.
[0036] Figure 4 A heatmap for analyzing the correlation between target sensors based on copula entropy;
[0037] Figure 5 The result of the Seq2Seq model reconstruction of the target sensor (using the normal operation data of the associated sensor combination associated with the target sensor as training data for reconstruction);
[0038] Figure 6 The result of the Seq2Seq model reconstruction of the target sensor (only the operating data when the target sensor is faulty is used as training data for reconstruction). Detailed Implementation
[0039] The present invention will now be further described with reference to the accompanying drawings.
[0040] The program code for this invention was compiled on an Intel® Core™ i7-4790 processor, 8GB of RAM, and NVIDIA GeForce RTX graphics. TM The system was built on a 2080Ti GPU and implemented in Python, using PyCharm Community Edition 2021.1.1 x64 as the integrated development platform. Its main function is to reconstruct data from the target sensor under fault conditions. The steam generator (A) feedwater flow rate (code WFWA) sensor was selected as the target sensor, and steady-state normal operation data and fault state data were acquired through the PCtran nuclear power plant simulation platform. First, copula entropy was used as the correlation metric to identify a set of sensor combinations associated with the target sensor under steady-state conditions. Based on this correlation structure, a Seq2Seq model was applied to train and test data reconstruction of the target sensor data under fault conditions. Figure 1 The specific steps of this invention are as follows:
[0041] Step 1: Collect historical sensor data of the system under steady-state normal operation and specific fault conditions using the PCtran nuclear power plant simulation platform.
[0042] Step 2: Using Copula entropy as a correlation metric, calculate the correlation strength between the target sensor and the time series data of all other sensors (except itself) in the dataset, and visualize all the calculated Copula entropy values as a heat map. Figure 4 In this model, darker colors (such as deep red) indicate a stronger correlation, while lighter colors (such as light blue) indicate a weaker correlation. By observing the heatmap and entropy values, a group of sensors with the strongest correlation to the target sensor is selected to form a correlated sensor combination.
[0043] The formula for calculating copula entropy is:
[0044] H c (X,Y)=-∫ u c(u,v)logc(u,v)dudv
[0045] Where, u = F X (x), v = F Y (y) is the marginal cumulative distribution function transformation of N-dimensional random variables X (target sensor) and Y (sensor data used for correlation calculation in the data).
[0046] Step 3: Data Preprocessing Stage
[0047] Sliding window construction: The system uses 10 time steps (from historical time t-9 to current time t) as a window to read 80-dimensional raw data (8 sensors × 10 time steps) from 8 associated sensors; Normalization processing: MinMax normalization (compressed to the range of 0-1) is performed on the 10-step time series data of each sensor to eliminate dimensional differences; 3D tensor generation: An input tensor with dimensions of [batch size, time step, number of features] (e.g., [32, 10, 8]) is generated.
[0048] Step 4: Training Phase: From the normal steady-state data, extract the readings of the above-mentioned associated sensor combination as input features, and extract the readings of the target sensor as the target output. After normalizing the extracted data, train the SeqSeq model. When the model training error converges to the preset accuracy or reaches the maximum number of iterations, stop training and save the optimal trained Seq2Seq model.
[0049] Step 5: Testing Phase: Extract readings from the fault data for the same associated sensor combinations and perform the same normalization process. Use these readings as input to the trained Seq2Seq model. Input the processed fault data into the trained Seq2Seq model, and the model will output the normal values of the target sensor under fault conditions (i.e., the reconstructed result).
[0050] Specifically, the encoder and decoder workflows of the Seq2Seq-based nuclear power plant sensor data reconstruction method are as follows:
[0051] Encoder workflow:
[0052] Bidirectional feature extraction: Forward LSTM layer: processes the time sequence from left to right, capturing the forward time dependency; Reverse LSTM layer: processes the time sequence from right to left, capturing the reverse time dependency; Each time step outputs a 64-dimensional hidden state (128-dimensional after bidirectional concatenation); Context memory generation: The hidden state at the final time step is used as a global context vector (128-dimensional), carrying the compressed information of the entire sequence.
[0053] Decoder workflow:
[0054] Initial state injection: The encoder's context vector is used as the initial state of the decoder LSTM; Stepwise prediction mechanism: Each prediction step receives the prediction value from the previous step (the first step uses the encoder's true value at the last moment); Combine the current LSTM state to generate a 1D output (target sensor prediction value), and adjust the output scale to match the original data distribution through a fully connected layer; Autoregressive loop: Repeat the prediction steps 10 times to construct a complete reconstruction sequence.
[0055] Example 1: Figure 2This is a diagram of the Seq2Seq model architecture, which automatically utilizes GPU acceleration when available. The model consists of an encoder-decoder structure with two LSTM layers in both components, each containing 64 hidden units. Regularization between layers uses a loss ratio of 0.1. The input dimension has 8 features depending on the sensor's relevance, while the output dimension is fixed at 1 for single-sensor value prediction. Temporal data preparation is performed using a 10-time-step sliding window method, where minimum scaling (0-1 range) is applied to both input and output variables. The model is trained using an Adam optimizer with a learning rate of 0.001 and L2 regularization (weight decay = 0.01). A batch size of 32 is used to automatically detect and utilize available GPU resources, while falling back to CPU computation when necessary. This architecture demonstrates particular effectiveness in handling the temporal dependencies of sensor data, while maintaining robustness through its regularization strategy.
[0056] Example 2: Figure 3 The following is a framework diagram of the nuclear power sensor data reconstruction method based on the Seq2Seq network model. The specific steps are as follows: (1) Steady-state normal operation data and fault data were obtained through the PCtran nuclear power plant simulation platform; (2) The steam generator (A) feedwater flow (code WFWA) sensor was selected as the target sensor; (3) Copula entropy was used as the correlation measurement method to identify the sensor combination (excluding the target sensor itself) associated with the target sensor under steady-state conditions; (4) Based on this correlation structure, the Seq2Seq model was trained using the associated sensor combination and target sensor data in the normal data; (5) The Seq2Seq model was reconstructed under fault conditions using the associated sensor combination and target sensor data in the fault data.
[0057] Example 3: Figure 4 A heatmap is used to analyze the correlation between target sensors based on copula entropy. Each row in the heatmap contains 10 data points, for a total of 9 rows. Since there are a total of 82 sensor columns, the cells in columns 3 to 10 of the ninth row are empty. Each element in the heatmap represents the Copula entropy (CE) value between the current target sensor and another sensor indexed by that column. For example, the values in the first row and second column represent the CE value between the steam generator (A) feedwater flow sensor and the sensor with column index 2 (RCS average temperature). A higher CE value indicates a stronger correlation between the two sensors, while a lower CE value indicates a weaker correlation.
[0058] Example 4: Figure 5The image shows the reconstruction results of the Seq2Seq model under four fault types of the target sensor (fixed bias, drift, accuracy degradation, and complete failure). The reconstruction data is based on normal operating data of the associated sensor combination used as training data. The horizontal axis represents the length of the time series sample, and the vertical axis represents the feedwater flow rate. The Seq2Seq model exhibits strong reconstruction performance under the four fault conditions of the target sensor (WFWA). Specifically, the reconstructed curves for different fault types of the target sensor show similar trends. The reconstruction results are consistent with normal data, demonstrating good robustness and high fidelity. Figure 6 The image shows the Seq2Seq model reconstruction results for four fault types of the target sensor (fixed bias, drift, accuracy degradation, and complete failure). (Only operational data during target sensor failures were used as training data for reconstruction.) The horizontal axis represents the length of the time series sample, and the vertical axis represents the water flow rate. Figure 5 Compared to the reconstruction results, the Seq2Seq model has a larger reconstruction error, exhibiting poor reconstruction performance, and the reconstruction results have a large error compared to the normal data.
[0059] In particular, in some preferred embodiments of the present invention, a computer device is also provided, including a memory and a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the data reconstruction method based on the Seq2Seq network model described in any of the above embodiments.
[0060] In some other preferred embodiments of the present invention, a computer-readable storage medium is also provided, on which a computer program / instruction is stored, wherein when the computer program is executed by a processor, it implements the steps of the data reconstruction method based on the Seq2Seq network model described in any of the above embodiments.
[0061] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the data reconstruction method embodiments based on the Seq2Seq network model described above, which will not be repeated here.
[0062] Computer-readable storage media encompass a variety of types, including persistent and non-persistent, portable and fixed. These media store information using different technologies, and the content can be machine instructions, data structures, program modules, or other types of data. Some typical examples of computer storage media include: phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), various types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory and other storage technologies, optical storage media such as CD-ROM and digital video disc (DVD), magnetic storage devices such as magnetic tape and disks, and other non-transferable media used to store information accessible to computing devices. It is important to note that the computer-readable media described herein do not include temporary storage media, such as modulated data signals and carrier waves.
[0063] Those skilled in the art will further recognize that the operation of the module can be achieved using existing technical protocols or programs, without relying on new computer programs themselves. The units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.
[0064] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented in hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.
[0065] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A data reconstruction method based on a Seq2Seq network model, characterized in that: Includes the following steps: Step 1: Obtain steady-state normal operation data and fault data through the nuclear power plant simulation platform, perform preprocessing, and select the target sensor; Step 2: Using copula entropy as a correlation measurement method, identify sensor combinations associated with the target sensor under steady-state conditions; Step 3: Based on the correlation structure in Step 2, a Seq2Seq model is trained and constructed using the associated sensor combination in the normal operation data and the target sensor data; Step 4: Utilize the associated sensor combination data in the fault data, and output the reconstructed data of the target sensor through the trained Seq2Seq model to realize the data reconstruction function of the target sensor under fault conditions.
2. The data reconstruction method based on the Seq2Seq network model according to claim 1, characterized in that: The sensor assembly does not include the target sensor itself.
3. The data reconstruction method based on the Seq2Seq network model according to claim 2, characterized in that: The formula for calculating the copula entropy in step 2 is as follows: H c (X,Y)=-∫ u c(u,v)logc(u,v)dudv Where, u = F X (x), v = F Y (y) is the marginal cumulative distribution function transformation of N-dimensional random variables X and Y, where X and Y are the sensor data used for correlation calculation in the target sensor and the dataset.
4. The data reconstruction method based on the Seq2Seq network model according to claim 3, characterized in that: After obtaining the copula entropy values of the target sensor and the other sensors, the correlation calculation results are presented in the form of a heat map. The darker the color, the stronger the correlation, and vice versa. Then, the sensor combination with the strongest correlation to the target sensor is selected by the color of the elements in the heat map and the magnitude of their entropy values. The normal operation data of the sensor combination with the strongest correlation is used as the input of the subsequent data reconstruction model.
5. The data reconstruction method based on a Seq2Seq network model according to claim 1, characterized in that: The training data for the Seq2Seq model has been preprocessed: The input data is constructed using a sliding window with 10 time steps, covering the period from historical time t-9 to the current time t. MinMax normalization is performed on the time series data of each sensor individually, compressing it to the range of [0,1]. Generate a three-dimensional input tensor with dimensions of [batch size, time step, number of features].
6. The data reconstruction method based on a Seq2Seq network model according to claim 5, characterized in that: The Seq2Seq model constructed in step 3 includes: Encoder: The input data is processed by a bidirectional LSTM layer. The forward LSTM layer captures the forward time dependency from left to right, and the backward LSTM layer captures the inverse time dependency from right to left. At each time step, the output is a bidirectionally concatenated 128-dimensional hidden state. The hidden state at the final time step is used as a 128-dimensional global context vector. Decoder: Using the encoder's context vector as the initial state, it generates the reconstructed sequence of the target sensor through a stepwise autoregressive prediction mechanism. At each step, it receives the prediction value from the previous step and outputs the 1D prediction result through a fully connected layer. The complete sequence is constructed by looping 10 times.
7. A data reconstruction method based on a Seq2Seq network model according to claim 1 or 6, characterized in that: The output data from step 4 is reconstructed and then post-processed. Inverse normalization: Maps the model's output predictions in the 0-1 range back to the original physical dimensions; Sliding window fusion: Overlapping window prediction results are fused using a weighted average to eliminate boundary effects; Final output: Generates target sensor reconstruction values that are fully synchronized with the original data sampling rate.
8. A data reconstruction method based on a Seq2Seq network model according to claim 6, characterized in that: The training termination condition for the Seq2Seq model is: reaching the preset maximum number of iterations or the mean square error between the model output and the theoretical output meets the convergence threshold.
9. A computer device / equipment / system, comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 8.
10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that: When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 8.